Gas power plant redundancy control method and system based on multi-modal perception and AI decision

Abstract

The invention relates to a gas power plant redundancy control method and system based on multi-modal perception and AI decision. The method comprises the following steps: receiving operation data of an execution mechanism; the execution mechanism operation data is multi-modal data collected by a sensor of the execution mechanism and preprocessed by an edge gateway; the execution mechanism is equipment responsible for executing instructions in the gas power plant; according to the operation data of the execution mechanism, calculating a fault prediction probability of the execution mechanism through an LSTM neural network; according to the operation data of the execution mechanism, calculating optimization parameters of the execution mechanism through a reinforcement learning algorithm; generating an executable strategy based on the fault prediction probability and the optimization parameter; sending the executable strategy to the execution mechanism; the executable strategy is used for indicating the execution mechanism to adjust own operation parameters. Through the steps, the effects of adaptively adjusting the operation parameters of the equipment and saving manpower and energy are achieved.

Description

Technical Field

[0001] The present invention belongs to the field of digital monitoring and control, and particularly relates to a redundant control method and system for a gas power plant based on multi-modal perception and AI decision-making. Background Art

[0002] With the development of industrial control technology, fieldbus control technology has emerged. By connecting all devices to the same fieldbus for reception and connecting field devices (such as sensors, actuators, controllers, etc.) through a unified digital network, it replaces the traditional point-to-point wiring method. Its core is to achieve data transmission and control instruction interaction through a shared communication line. However, the traditional fieldbus control technology has problems such as the dependence of device operation parameters on manual experience adjustment and the possible paralysis of the entire network caused by a single-point failure of the bus. Summary of the Invention

[0003] Based on this, in view of the above technical problems, it is necessary to provide a redundant control method for a gas power plant based on multi-modal perception and AI decision-making that can adaptively adjust device operation parameters and save manpower and energy.

[0004] In a first aspect, the present application provides a redundant control method for a gas power plant based on multi-modal perception and AI decision-making, including:

[0005] Receiving the operation data of the actuator; the operation data of the actuator is multi-modal data collected by the sensors of the actuator and preprocessed by the edge gateway; the actuator is a device responsible for executing instructions in the gas power plant;

[0006] Calculating the fault prediction probability of the actuator through an LSTM neural network according to the operation data of the actuator;

[0007] Calculating the optimization parameters of the actuator through a reinforcement learning algorithm according to the operation data of the actuator;

[0008] Generating an executable policy based on the fault prediction probability and the optimization parameters;

[0009] Sending the executable policy to the actuator; the executable policy is used to instruct the actuator to adjust its own operation parameters.

[0010] Further, calculating the fault prediction probability of the actuator through an LSTM neural network according to the operation data of the actuator includes:

[0011] Inputting the operation data of the actuator into a bidirectional LSTM layer to extract temporal features and obtaining a feature vector;

[0012] Among them, the bidirectional LSTM layer extracts temporal features through the following formula:

[0013]

[0014] Among them, is the forward hidden state, is the backward hidden state, and g(·) is the fusion function;

[0015] The feature vector is segmented into sub - windows to obtain window feature vectors;

[0016] The window feature vectors are input into the convolutional layer for convolution to obtain the spatio - temporally focused feature matrix;

[0017] The key features of the feature matrix are dynamically weighted to obtain the weighted feature matrix;

[0018] The weighted feature matrix is input into the prediction layer to obtain the prediction result;

[0019] The prediction results are fused to obtain the fault prediction probability of the actuator.

[0020] Furthermore, the prediction layer obtains the prediction result through the following method:

[0021] The weighted feature matrix is input into the global average pooling layer to obtain the pooled feature matrix;

[0022] The pooled feature matrix is dimension - reduced through a fully - connected layer to obtain the dimension - reduced feature;

[0023] The dimension - reduced feature is activated to obtain the fault probability scalar; the fault probability scalar represents the fault probability of each actuator;

[0024] The weighted feature matrix is input into the GRU layer to extract the medium - term temporal features to obtain the medium - term features;

[0025] The extracted features are input into the time attention layer to obtain the time - window features; and the time - window features are linearly regressed to obtain the daily performance degradation rate of the actuator;

[0026] The weighted feature matrix is input into the Weibull model to obtain the Weibull parameters;

[0027] The Weibull parameters are input into the remaining life calculation layer to obtain the remaining life of the actuator;

[0028] Among them, the loss function is calculated by the following formula:

[0029]

[0030] Among them, is the improved Focal Loss function, is the Huber Loss function, is the Weibull negative log - likelihood;

[0031] Among them, the weight is calculated by the following formula:

[0032]

[0033] Among them, σ i is a trainable parameter;

[0034] The backpropagation loss function is used to update the prediction layer;

[0035] Based on the failure probability scalar, the daily performance degradation rate, and the remaining life, the prediction result is obtained.

[0036] Furthermore, the Weibull model obtains the Weibull parameters through the following method:

[0037] The weighted feature matrix is split into multiple heads to obtain a multi-head feature matrix;

[0038] Each multi-head feature matrix is independently used to calculate the QKV matrix to obtain an attention subspace;

[0039] The features of the attention subspace are fused to obtain a long-term feature matrix;

[0040] Based on the long-term feature matrix, the original Weibull parameters are calculated through the following formula:

[0041] λ, k = W λ ·h attn + b λ , W k ·h attn + b k

[0042] Among them, λ is the scale parameter that controls the change of life, and k is the shape parameter that controls the change trend of the failure rate;

[0043] The original Weibull parameters are processed by the following formula to obtain the Weibull parameters:

[0044] λ ∈ (0, +∞)

[0045] k ∈ (1, +∞)

[0046] Among them, k = 1 represents a constant failure rate for random failures, and k > 1 represents an increasing failure rate for aging failures.

[0047] Furthermore, after receiving the operating data of the actuator, it also includes:

[0048] The three-dimensional BIM model coordinate system of the actuator is matched and fused with the actual geographical coordinate system to obtain a coordinate matching model;

[0049] Assign a unique ID to each component in the coordinate matching model. This ID corresponds one-to-one with the actuator to obtain the ID mapping relationship table;

[0050] Based on the ID mapping relationship table, associate the actuator operation data and the fault prediction probability with the corresponding components in the 3D BIM model to obtain a 3D model with dynamic data binding;

[0051] Send the 3D model with dynamic data binding to the terminal; this 3D model with dynamic data binding is used for display on the terminal display device;

[0052] Among them, the terminal is used to feedback virtual instructions; the virtual instructions are used to instruct the actuator to change its own operating state.

[0053] Furthermore, after sending the 3D model with dynamic data binding to the terminal, it further includes:

[0054] Receive virtual device instructions; these virtual device instructions are feedback by the terminal;

[0055] Generate a virtual device based on the virtual device instructions and the BIM model;

[0056] Connect the virtual device to the 3D model with dynamic data binding to obtain a new 3D model;

[0057] Execute the control strategy, drive the new 3D model to act, and calculate the dynamic response of the new 3D model;

[0058] Based on the dynamic response, verify the logic compliance and generate a virtual device result prediction;

[0059] Generate a virtual commissioning report based on the virtual device result prediction.

[0060] Furthermore, after sending the 3D model with dynamic data binding to the terminal, it further includes:

[0061] Receive preset scenario instructions; these preset scenario instructions are feedback by the terminal;

[0062] Inject faults into the 3D model with dynamic data binding based on the predicted scenario to obtain a 3D model with faults;

[0063] Based on the 3D model with faults, predict the fault impact through the AI model to generate a fault impact prediction;

[0064] Generate fault handling steps based on the fault impact prediction;

[0065] Generate an optimal emergency plan based on the fault handling steps.

[0066] Furthermore, this method further includes:

[0067] It is detected that the operation data of the actuator has not been received through the fieldbus for more than 200 ms, and a bus exception signal is generated;

[0068] Based on the bus exception signal, the executable policy is asymmetrically encrypted using the public key of the actuator to obtain the encrypted policy;

[0069] Calculate the hash of the encrypted policy to obtain the encrypted policy hash value;

[0070] Broadcast the encrypted policy hash value and the encrypted policy to the blockchain network to obtain the blockchain transaction hash;

[0071] Send the encrypted policy and the blockchain transaction hash to the target actuator through the 5G network, for instructing the target actuator to feedback the hash verification result through the 5G network;

[0072] Among them, the blockchain network permanently stores the encrypted policy hash value and the encrypted policy.

[0073] Furthermore, according to the operation data of the actuator, optimize the parameters through the reinforcement learning algorithm, including:

[0074] Normalize the electricity price curve, load demand and actuator operation data to obtain the normalized data;

[0075] Encode the normalized data into a state vector;

[0076] Input the state vector into the deep deterministic policy gradient model and output a continuous action vector;

[0077] Perform constraint processing on the continuous action vector to obtain the optimized parameters;

[0078] Among them, the deep deterministic policy gradient model calculates the continuous action vector through the following formula:

[0079]

[0080] Q(s t ,a t )=r t +γ·Q'(s t+1 ,μ'(s t+1 ))

[0081] Among them, is the gradient of the policy objective function with respect to the policy network parameters, s i is the state of the i-th sample, μ(s i ) is the action generated by the policy network according to the state, Q(s t ,a t ) is the Q-value prediction of the current state, r tis the immediate reward at the current moment, and γ is the discount factor;

[0082] Among them, the model is updated through the following formula:

[0083] θ Q' ← τθ Q +(1 - τ)θ Q’ , θ μ' ← τθ μ +(1 - τ)θ μ

[0084] Among them, τ is the update rate.

[0085] In a second aspect, the present application further provides a redundant control system for a gas power plant based on multi-modal perception and AI decision-making, including:

[0086] A receiving module, configured to receive the operation data of the actuator; the operation data of the actuator is multi-modal data collected by the sensors of the actuator and preprocessed by the edge gateway; the actuator is a device in the gas power plant responsible for executing instructions;

[0087] A fault prediction module, configured to calculate the fault prediction probability of the actuator through an LSTM neural network according to the operation data of the actuator;

[0088] An optimization parameter module, configured to calculate the optimization parameters of the actuator through a reinforcement learning algorithm according to the operation data of the actuator;

[0089] A decision-making module, configured to generate an executable policy based on the fault prediction probability and the optimization parameters;

[0090] An adjustment module, configured to send the executable policy to the actuator; the executable policy is used to instruct the actuator to adjust its own operation parameters.

[0091] The above-mentioned redundant control method and system for a gas power plant based on multi-modal perception and AI decision-making receive the operation data of the actuator; the operation data of the actuator is multi-modal data collected by the sensors of the actuator and preprocessed by the edge gateway; the actuator is a device in the gas power plant responsible for executing instructions; calculate the fault prediction probability of the actuator through an LSTM neural network according to the operation data of the actuator; calculate the optimization parameters of the actuator through a reinforcement learning algorithm according to the operation data of the actuator; generate an executable policy based on the fault prediction probability and the optimization parameters; send the executable policy to the actuator; the executable policy is used to instruct the actuator to adjust its own operation parameters. Through the above steps, it is possible to solve the problems that the existing traditional fieldbus control technology depends on manual experience to adjust the operation parameters of equipment, and a single-point bus fault may cause the entire network to collapse, and achieve the effects of being able to adaptively adjust the operation parameters of equipment, reducing the impact of bus faults, and saving manpower and energy. Brief Description of the Drawings

[0092] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the following will briefly introduce the drawings required for use in the description of the embodiments or related technologies. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0093] Figure 1 It is a flowchart of the redundant control method for a gas power plant based on multi-modal perception and AI decision-making of the present invention;

[0094] Figure 2 It is a system diagram of the redundant control system for a gas power plant based on multi-modal perception and AI decision-making of the present invention; Detailed Embodiments

[0095] In order to make the objectives, technical solutions and advantages of the present application more clearly understood, the following further details the present application in conjunction with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0096] The redundant control method for a gas power plant based on multi-modal perception and AI decision-making provided by the embodiments of the present application can be applied to the application environment of digital gas power plant control.

[0097] In one embodiment, as Figure 1 shown, a redundant control method for a gas power plant based on multi-modal perception and AI decision-making is provided. In this embodiment, it is exemplified that the method is applied to a central server. It can be understood that the method can also be applied to a terminal, and can also be applied to a system including a terminal and a server, and is realized through the interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0098] Step 101, receiving the operation data of the actuator.

[0099] Specifically, the operation data of the actuator is multi-modal data collected by the sensors of the actuator and preprocessed by the edge gateway. The actuator is a device in the gas power plant responsible for executing instructions. The edge gateway is a computing node deployed locally on the device, responsible for data preprocessing, filtering redundant information, and reducing the transmission load to the cloud. Obtain the real-time operation data from the actuators in the gas power plant. These data are collected by a variety of sensors and preliminarily processed by the edge gateway, and finally form multi-modal data.

[0100] Exemplarily, the actuator may include valves, turbines, pumps, etc., and the sensors may include temperature sensors, pressure sensors, vibration sensors, etc.

[0101] Step 102: Calculate the fault prediction probability of the actuator based on the actuator operation data through an LSTM neural network.

[0102] Specifically, analyze the time-series operation data (such as vibration and temperature change trends) of the actuator through a bidirectional LSTM layer to predict the probability of the device failing in the future. The fault prediction probability is used to quantify the likelihood of the device failing within a certain period in the future. A bidirectional LSTM is an improved recurrent neural network structure composed of two LSTM layers, a forward layer and a backward layer. The forward layer processes the sequence data in chronological order to capture historical information, and the backward layer processes the data in reverse order to capture potential future trends.

[0103] Step 103: Calculate the optimization parameters of the actuator based on the actuator operation data through a reinforcement learning algorithm.

[0104] Specifically, simulate the states (such as energy consumption and efficiency) of the actuator under different operating parameters through Deep Deterministic Policy Gradient (DDPG) to dynamically calculate the optimal parameter combination (such as adjusting the valve opening and turbine speed) to improve performance or reduce energy consumption.

[0105] Step 104: Generate an executable policy based on the fault prediction probability and the optimization parameters.

[0106] Specifically, the executable policy is an instruction or operation suggestion that the device can directly execute, usually including parameter adjustment instructions and maintenance plans. Combine reliability constraints and performance goals to generate a policy that takes into account both safety and efficiency.

[0107] Step 105: Send the executable policy to the actuator.

[0108] Specifically, the executable policy is used to instruct the actuator to adjust its own operating parameters. Send the executable policy to the actuator through a fieldbus and / or 5G network channel, and the device automatically adjusts the operating parameters. At the same time, the actuator returns the operation data after adjusting the operating parameters.

[0109] The redundant control method for gas power plants based on multi-modal perception and AI decision-making provided by the embodiments of the present application receives the operation data of the actuator; the operation data of the actuator is multi-modal data collected by the sensors of the actuator and pre-processed by the edge gateway; the actuator is the device responsible for executing instructions in the gas power plant; according to the operation data of the actuator, the fault prediction probability of the actuator is calculated through the LSTM neural network; according to the operation data of the actuator, the optimization parameters of the actuator are calculated through the reinforcement learning algorithm; based on the fault prediction probability and the optimization parameters, an executable policy is generated; the executable policy is sent to the actuator; the executable policy is used to indicate the steps for the actuator to adjust its own operation parameters, which can solve the problems that the traditional fieldbus control technology relies on manual experience to adjust the device operation parameters and the single-point failure of the bus may lead to the paralysis of the entire network, and realizes the effects of being able to adaptively adjust the device operation parameters, reducing the impact of bus failures, and saving manpower and energy.

[0110] In one of the embodiments, according to the operation data of the actuator, calculating the fault prediction probability of the actuator through the LSTM neural network includes:

[0111] Step 201, input the operation data of the actuator into the bidirectional LSTM layer to extract temporal features, and obtain a feature vector.

[0112] Among them, the bidirectional LSTM layer extracts temporal features through the following formula:

[0113]

[0114] Specifically, is the forward hidden state. is the backward hidden state. g(·) is a fusion function, an operation to merge bidirectional information, and common methods include concatenation or weighted summation. Input the multi-modal temporal data of the actuator into the bidirectional LSTM layer, extract the hidden states through the forward LSTM and the backward LSTM respectively, and then merge the bidirectional information through the fusion function to form a vector containing the complete context temporal features.

[0115] Step 202, perform sub-window segmentation on the feature vector to obtain window feature vectors.

[0116] Specifically, the temporal feature vector is segmented into multiple sub-windows according to a fixed time length, and the data within each sub-window is used to capture local temporal patterns. Sub-window segmentation cuts long temporal data into short segments to facilitate subsequent focusing on local features. The window feature vector is the set of features at all time points within the sub-window.

[0117] Step 203, input the window feature vector into the convolutional layer for convolution to obtain a spatio-temporally focused feature matrix.

[0118] Specifically, the window features are input into the convolutional layer. By sliding the convolutional kernel in the time and feature dimensions, local spatio-temporal features are extracted. Spatio-temporal focusing captures the correlations between adjacent time points and feature dimensions simultaneously during convolution.

[0119] Step 204: Dynamically weight the key features of the feature matrix to obtain a weighted feature matrix.

[0120] Specifically, calculate the dynamic weights for each spatio-temporal position in the feature matrix to enhance important features and suppress noise. The dynamic weights adaptively assign weights according to feature importance, such as using attention scores or gating mechanisms.

[0121] Step 205: Input the weighted feature matrix into the prediction layer to obtain the prediction result.

[0122] Specifically, input the weighted feature matrix into the prediction layer model to obtain the short-term, medium-term, and long-term fault probabilities, and integrate them to obtain the prediction result.

[0123] Step 206: Fuse the prediction results to obtain the fault prediction probability of the actuator.

[0124] Specifically, aggregate all the prediction results to generate the final fault prediction probability for each actuator.

[0125] In this embodiment, by introducing a double-layer LSTM layer to extract features, the recognition ability for complex scenarios is enhanced, and the accuracy of predicting fault situations is improved.

[0126] In one of the embodiments, the prediction layer obtains the prediction result through the following method:

[0127] Step 301: Input the weighted feature matrix into the global average pooling layer to obtain a pooled feature matrix.

[0128] Specifically, input the weighted feature matrix into the global average pooling layer, take the average value of each feature channel over all time steps, compress it into a pooled feature matrix, eliminate the redundancy in the time dimension, and retain the core features.

[0129] Step 302: Reduce the dimension of the pooled feature matrix through a fully connected layer to obtain the reduced-dimensional features.

[0130] Specifically, input the pooled features into the fully connected layer, map the high-dimensional features to a low-dimensional space through linear transformation, reduce the computational amount, and extract high-level abstract features.

[0131] Step 303: Activate the reduced-dimensional features to obtain a fault probability scalar; the fault probability scalar represents the fault probability of each actuator.

[0132] Specifically, an activation function (such as Sigmoid) is applied to the features after dimensionality reduction, and a scalar value in the range of 0 to 1 is output, representing the failure probability of the current state of the actuator.

[0133] Step 304: Input the weighted feature matrix into the GRU layer to extract medium-term time series features, obtaining medium-term features.

[0134] Specifically, the weighted feature matrix is input into the GRU (Gated Recurrent Unit) layer to capture medium-term dependencies in the time series. GRU is a simplified version of the recurrent neural network that alleviates the vanishing gradient problem through a gating mechanism and is suitable for medium-length time series modeling.

[0135] Step 305: Input the extracted features into the time attention layer to obtain time window features; and perform linear regression on the time window features to obtain the daily performance degradation rate of the actuator.

[0136] Specifically, the time attention layer is used to dynamically assign weights to different time steps according to feature importance, outputting weighted time window features. Linear regression maps the time window features to the daily performance degradation rate.

[0137] Step 306: Input the weighted feature matrix into the Weibull model to obtain Weibull parameters.

[0138] Specifically, the weighted feature matrix is input into the Weibull distribution parameter estimation layer, outputting the shape parameter and the scale parameter, which describe the device failure law.

[0139] Step 307: Input the Weibull parameters into the remaining life calculation layer to obtain the remaining life of the actuator.

[0140] Specifically, according to the Weibull parameters, the Weibull survival function is used to calculate the remaining service life of the device in the current state.

[0141] Step 308: Backpropagate the loss function to update the prediction layer.

[0142] Among them, the loss function is calculated through the following formula:

[0143]

[0144] Among them, the weights are calculated through the following formula:

[0145]

[0146] Specifically, The improved Focal Loss function is used to solve the problem of unbalanced fault categories. The Huber Loss loss function is a robust regression loss used to reduce the impact of outliers on the prediction of the performance degradation rate. The Weibull negative log-likelihood is to maximize the likelihood probability of the observed lifetime data. Through the trainable parameter σ i Adaptively adjust the weights of each task.

[0147] Step 309, obtain the prediction result based on the failure probability scalar, the daily performance degradation rate, and the remaining life.

[0148] Specifically, combine the failure probability scalar (short-term risk), the daily performance degradation rate (mid-term trend), and the remaining life (long-term reliability) to output a multi-dimensional prediction result to support maintenance decisions.

[0149] This embodiment improves the comprehensiveness and accuracy of failure prediction through multi-task collaborative joint optimization.

[0150] In one of the embodiments, the Weibull model obtains the Weibull parameters through the following method:

[0151] Step 401, split the weighted feature matrix into multiple heads to obtain a multi-head feature matrix.

[0152] Specifically, evenly divide the weighted feature matrix along the feature dimension into multiple sub-matrices, and each sub-matrix is responsible for capturing feature patterns from different angles.

[0153] Step 402, independently calculate the QKV matrix for each multi-head feature matrix to obtain the attention subspace.

[0154] Specifically, for the feature matrix of each head, generate the query matrix (Q), the key matrix (K), and the value matrix (V) respectively through linear transformation, calculate the correlation weights at different time steps through the self-attention mechanism, and generate the attention subspace features. The QKV matrix is used to measure the correlation between features. The attention subspace is used to reflect the dependence relationship at different time steps.

[0155] Step 403, fuse the attention subspace features to obtain the long-term feature matrix.

[0156] Specifically, splice or weight-fuse the attention subspace features of all heads to form a comprehensive long-term feature matrix to capture the cross-time dependence in the device degradation process. The long-term feature matrix is used to describe the overall trend of the device performance degradation over time.

[0157] Step 404, based on the long-term feature matrix, calculate the original Weibull parameters through the following formula:

[0158] λ, k = W λ ·h attn +bλ , W k ·h attn +b k

[0159] Specifically, the long-term feature matrix is input into the fully connected layer, and the scale parameter and shape parameter of the Weibull distribution are respectively output through linear transformation. λ is the scale parameter, which is used to control the life variation. k is the shape parameter, which is used to control the change trend of the failure rate.

[0160] Step 405, perform constraint processing on the original Weibull parameters through the following formula to obtain the Weibull parameters:

[0161] λ ∈ (0, +∞)

[0162] k ∈ (1, +∞)

[0163] Specifically, the value range of the parameters is restricted to conform to the definition of the Weibull distribution to ensure reasonable model output. λ is transformed through the exponential function, and k is determined by adding 1 to the ReLU activation function. k = 1 represents a constant failure rate, which is applicable to the sudden failure of electronic components. k > 1 represents an increasing failure rate, which is applicable to mechanical wear.

[0164] In this embodiment, λ and k are adjusted according to real-time data to achieve personalized prediction of the remaining life, increasing the accuracy of fault prediction.

[0165] In one of the embodiments, after receiving the operation data of the actuator, it further includes:

[0166] Step 501, match and fuse the three-dimensional BIM model coordinate system of the actuator with the actual geographical coordinate system to obtain a coordinate matching model.

[0167] Specifically, the actual geographical coordinates are obtained through GPS positioning and on-site surveying and mapping. The three-dimensional BIM model (Building Information Model) of the actuator is aligned with the actual geographical coordinates to ensure that the virtual model is consistent with the actual physical position, forming a coordinate matching model. The position, rotation, and scaling parameters of the BIM model are adjusted through the coordinate transformation algorithm. The geographical coordinate system is the physical space coordinate based on the earth's longitude and latitude or local surveying and mapping. The coordinate matching model is a virtual model that is completely aligned with the actual space and is used for real-scene mapping.

[0168] Step 502, assign a unique ID to each component in the coordinate matching model. This ID corresponds one-to-one with the actuator to obtain an ID mapping relationship table.

[0169] Specifically, the unique ID is a globally unique identifier used to distinguish different components.

[0170] Step 503: Based on the ID mapping relation table, associate the operation data of the actuator and the fault prediction probability with the corresponding components of the 3D BIM model to obtain a 3D model with dynamic data binding.

[0171] Specifically, according to the ID mapping table, associate the real-time operation data (such as temperature and vibration) and the fault prediction probability with the corresponding components of the BIM model through the API or message queue, so that the virtual model dynamically displays the actual state. The virtual model is linked with the real-time data to achieve state visualization.

[0172] Exemplarily, when the vibration data associated with a certain valve ID exceeds the threshold, the corresponding component of the model displays a red warning.

[0173] Step 504: Send the 3D model with dynamic data binding to the terminal; this 3D model with dynamic data binding is used for display on the display device of the terminal.

[0174] Specifically, the terminal is used to feedback virtual instructions. The virtual instructions are used to instruct the actuator to change its own operating state. Send the dynamically bound 3D model to the terminal (such as an industrial tablet or AR glasses) for the operator to view; the terminal supports sending virtual instructions, and drives the actuator to change the operating state through the control system.

[0175] In this embodiment, visual monitoring is realized through virtual-real mapping, forming a closed loop of "monitoring to decision-making to control", intuitively displaying the device state, simplifying the control process, and enhancing the user experience.

[0176] In one of the embodiments, after sending the 3D model with dynamic data binding to the terminal, it further includes:

[0177] Step 601: Receive virtual device instructions.

[0178] Specifically, obtain the virtual device instructions issued by the user from the terminal, and this instruction indicates adding or modifying virtual devices in the 3D model.

[0179] Step 602: Generate a virtual device based on the virtual device instructions and the BIM model.

[0180] Specifically, based on the component attributes in the BIM model and the virtual device instructions, dynamically generate new virtual devices in the 3D model and endow them with functional logics.

[0181] Step 603: Connect the virtual device to the 3D model with dynamic data binding to obtain a new 3D model.

[0182] Specifically, through the API or message middleware, logically bind the newly generated virtual device to the 3D model with dynamic data binding to form an interactive new 3D model.

[0183] Step 604: Execute the control strategy to drive the new 3D model to act and calculate the dynamic response of the new 3D model.

[0184] Specifically, drive the virtual device in the new 3D model according to the control strategy, and simulate the system behavior after the virtual device acts through the simulation engine.

[0185] Step 605: Verify the logical compliance based on the dynamic response and generate a prediction of the virtual device result.

[0186] Specifically, verify whether the dynamic response complies with the safety specifications and generate a prediction report based on the simulation data.

[0187] Step 606: Generate a virtual commissioning report based on the prediction of the virtual device result.

[0188] Specifically, summarize the simulation results, compliance verification conclusions, and optimization suggestions to form a virtual commissioning report to guide the actual system transformation or parameter adjustment.

[0189] In this embodiment, by introducing a virtual device, the impact of the newly added device is tested in a virtual environment to avoid the risks of physical experiments, and problems are discovered in advance through virtual commissioning, reducing the on-site trial-and-error time and resource consumption.

[0190] In one of the embodiments, after sending the 3D model with dynamic data binding to the terminal, it further includes:

[0191] Step 701: Receive a preset scenario instruction.

[0192] Specifically, the preset scenario instruction is fed back by the terminal. Receive the fault simulation scenario instruction predefined by the user, and this instruction specifies parameters such as the type, location, and severity of the fault to be injected.

[0193] Step 702: Inject a fault into the 3D model with dynamic data binding based on the predicted scenario to obtain a faulty 3D model.

[0194] Specifically, according to the preset scenario instruction, artificially introduce a fault state into the 3D model with dynamic data binding, such as modifying component parameters (e.g., the output of the temperature sensor is constantly 100 °C), simulating physical damage (e.g., a pipeline crack model), or data perturbation (e.g., adding noise to the vibration signal).

[0195] Step 703: Based on the faulty 3D model, predict the fault impact through an AI model and generate a fault impact prediction.

[0196] Specifically, input the faulty 3D model into a multimodal AI model (such as a model combining a graph neural network and time series prediction) to predict the impact of the fault on device performance, subsystem association, and production safety.

[0197] Step 704: Generate fault handling steps based on fault impact prediction.

[0198] Specifically, based on the AI prediction results, generate a specific operation process through a decision tree or a rule engine.

[0199] Step 705: Generate an optimal emergency plan based on the fault handling steps.

[0200] Specifically, comprehensively considering the fault impact prediction, available resources, and constraints, generate a multi-objective optimal emergency plan through an optimization algorithm.

[0201] In this embodiment, through fault simulation, the optimal solution is pre-calculated, reducing the possibility of major accidents, minimizing the losses of minor accidents, and enhancing stability.

[0202] In one of the embodiments, the method further includes:

[0203] Step 801: When it is detected that the operation data of the actuator has not been received through the fieldbus for more than 200 ms, generate a bus anomaly signal.

[0204] Specifically, monitor the communication status of the fieldbus in real time. If it is detected that the operation data of the actuator has not been received for 200 consecutive milliseconds, trigger a bus anomaly signal, indicating that the communication link may be interrupted or the delay is too high.

[0205] Step 802: Based on the bus anomaly signal, use the public key of the actuator to perform asymmetric encryption on the executable policy to obtain the encrypted policy.

[0206] Specifically, use the public key of the target actuator to encrypt the executable policy to be issued, generating an encrypted policy in ciphertext form to ensure that only the actuator holding the corresponding private key can decrypt it.

[0207] Step 803: Calculate the hash of the encrypted policy to obtain the encrypted policy hash value.

[0208] Specifically, calculate the hash value of a fixed length for the encrypted policy through a hash function for subsequent integrity verification. The hash value is the unique data fingerprint generated by the hash function, and any data tampering will cause the hash value to change.

[0209] Step 804: Broadcast the encrypted policy hash value and the encrypted policy to the blockchain network to obtain the blockchain transaction hash.

[0210] Specifically, broadcast the encrypted policy and its hash value to the blockchain network, and write the data into the block through the consensus mechanism to generate a unique blockchain transaction hash, achieving data immutability and permanent evidence storage.

[0211] Step 805: Send the encrypted policy and the blockchain transaction hash to the target executing agency via the 5G network, for instructing the target executing agency to feedback the hash verification result via the 5G network.

[0212] Specifically, send the encrypted policy and the transaction hash to the target executing agency via the 5G network, require the executing agency to decrypt and recalculate the hash value, compare it with the hash value stored in the blockchain, and feedback the verification result.

[0213] In this embodiment, by adding a 5G network data transmission channel, the risk of network paralysis caused by bus failures is reduced, and the stability is improved.

[0214] In one embodiment, according to the operating data of the executing agency, optimize the parameters by calculating with a reinforcement learning algorithm, including:

[0215] Step 901: Normalize the electricity price curve, load demand and the operating data of the executing agency to obtain normalized data.

[0216] Specifically, uniformly scale heterogeneous data such as time series electricity prices, electricity consumption, and the operating data of the executing agency to the same numerical range, eliminate the dimension difference, and improve the stability of model training.

[0217] Step 902: Encode the normalized data into a state vector.

[0218] Step 903: Input the state vector into the deep deterministic policy gradient model to output a continuous action vector.

[0219] Among them, the deep deterministic policy gradient model calculates the continuous action vector through the following formula:

[0220]

[0221] Q(s t ,a t )=r t +γ·Q'(s t+1 ,μ'(s t+1 ))

[0222] Specifically, is the gradient of the policy objective function with respect to the policy network parameters, s i is the state of the i-th sample, μ(s i ) is the action generated by the policy network according to the state, Q(s t ,a t ) is the Q-value prediction of the current state, r t is the immediate reward at the current moment, and γ is the discount factor.

[0223] Among them, the model is updated through the following formula:

[0224] θ Q' ← τθ Q + (1 - τ)θ Q’ , θ μ' ← τθ μ + (1 - τ)θ μ’

[0225] Specifically, τ is the update rate.

[0226] Step 904: Perform constraint processing on the continuous action vector to obtain optimization parameters.

[0227] Specifically, apply physical constraints (such as the upper limit of device power and the opening range of valves) to the output action vector to ensure that the actions are feasible in the actual system.

[0228] To further illustrate the solution of the embodiments of the present application, a specific example is given below for explanation.

[0229] 1. Hardware deployment example:

[0230] Install a multimodal sensor (sampling frequency 1 kHz) of model SEN - 2024 on the actuator of the circulating water pump in a gas power plant, and connect it to the edge gateway (NVIDIA Jetson AGX edge computing module) through the RS - 485 interface.

[0231] The 5G communication module uses the Huawei MH5000 - 31 industrial - grade module, which supports NSA / SA dual - mode and has a time delay ≤ 20 ms.

[0232] 2. Software implementation example:

[0233] Fault prediction model training:

[0234] Input data: The equipment failure event dataset in the past 3 years (including more than 1000 - dimensional features such as vibration spectrum and current waveform).

[0235] Output result: The accuracy rate of predicting mechanical failures 48 hours in advance is ≥ 92%.

[0236] Blockchain network construction:

[0237] Build a private chain based on Hyperledger Fabric, and generate a unique hash value for each control instruction and store it on the chain for evidence.

[0238] It should be understood that although the steps in the flowcharts involved in the above-described embodiments are sequentially shown according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless there is a clear description in this article, the execution of these steps has no strict order limit, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above-described embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be executed alternately or in turn with at least a part of other steps or steps or stages in other steps.

[0239] Based on the same inventive concept, an embodiment of the present application also provides a multi-modal perception and AI decision-based redundant control system for a gas power plant for implementing the above-mentioned multi-modal perception and AI decision-based redundant control method for a gas power plant. The implementation solutions provided by this system to solve problems are similar to the implementation solutions described in the above method. Therefore, the specific limitations in one or more embodiments of the multi-modal perception and AI decision-based redundant control system for a gas power plant provided below can refer to the limitations on the multi-modal perception and AI decision-based redundant control method for a gas power plant in the above text, and will not be repeated here.

[0240] In an exemplary embodiment, as Figure 2 shown, a multi-modal perception and AI decision-based redundant control system 1000 for a gas power plant is provided, including:

[0241] A receiving module 1001, configured to receive actuator operation data; the actuator operation data is multi-modal data collected by sensors of the actuator and preprocessed by an edge gateway; the actuator is a device in the gas power plant responsible for executing instructions;

[0242] A fault prediction module 1002, configured to calculate the fault prediction probability of the actuator through an LSTM neural network according to the actuator operation data;

[0243] An optimization parameter module 1003, configured to calculate the optimization parameters of the actuator through a reinforcement learning algorithm according to the actuator operation data;

[0244] A decision module 1004, configured to generate an executable policy based on the fault prediction probability and the optimization parameters;

[0245] An adjustment module 1005, configured to send the executable policy to the actuator; the executable policy is used to instruct the actuator to adjust its own operating parameters.

[0246] Furthermore, the fault prediction module is further configured to:

[0247] Input the operating data of the actuator into the bidirectional LSTM layer to extract temporal features and obtain a feature vector;

[0248] Among them, the bidirectional LSTM layer extracts temporal features through the following formula:

[0249]

[0250] Among them, is the forward hidden state, is the backward hidden state, and g(·) is the fusion function;

[0251] Perform sub-window segmentation on the feature vector to obtain window feature vectors;

[0252] Input the window feature vectors into the convolutional layer for convolution to obtain a spatio-temporally focused feature matrix;

[0253] Dynamically weight the key features of the feature matrix to obtain a weighted feature matrix;

[0254] Input the weighted feature matrix into the prediction layer to obtain a prediction result;

[0255] Fuse the prediction results to obtain the fault prediction probability of the actuator.

[0256] Furthermore, the prediction layer obtains the prediction result through the following method:

[0257] Input the weighted feature matrix into the global average pooling layer to obtain a pooled feature matrix;

[0258] Reduce the dimension of the pooled feature matrix through a fully connected layer to obtain a reduced-dimensional feature;

[0259] Activate the reduced-dimensional feature to obtain a fault probability scalar; the fault probability scalar represents the fault probability of each actuator;

[0260] Input the weighted feature matrix into the GRU layer to extract medium-term temporal features and obtain medium-term features;

[0261] Input the extracted features into the time attention layer to obtain time window features; and perform linear regression on the time window features to obtain the daily performance degradation rate of the actuator;

[0262] Input the weighted feature matrix into the Weibull model to obtain Weibull parameters;

[0263] Input the Weibull parameters into the remaining life calculation layer to obtain the remaining life of the actuator;

[0264] Among them, the loss function is calculated through the following formula:

[0265]

[0266] Among them, is the improved Focal Loss function, is the Huber Loss function, is the Weibull negative log-likelihood;

[0267] Among them, the weights are calculated by the following formula:

[0268]

[0269] Among them, σ i is a trainable parameter;

[0270] The backpropagation loss function updates the prediction layer;

[0271] Based on the failure probability scalar, the daily performance degradation rate, and the remaining life, the prediction result is obtained.

[0272] Furthermore, the Weibull model obtains the Weibull parameters through the following method:

[0273] The weighted feature matrix is split into multiple heads to obtain a multi-head feature matrix;

[0274] Each multi-head feature matrix independently calculates the QKV matrix to obtain an attention subspace;

[0275] The features of the attention subspace are fused to obtain a long-term feature matrix;

[0276] Based on the long-term feature matrix, the original Weibull parameters are calculated by the following formula:

[0277] λ, k = W λ ·h attn +b λ , W k ·h attn +b k

[0278] Among them, λ is the scale parameter that controls the life change, and k is the shape parameter that controls the change trend of the failure rate;

[0279] The original Weibull parameters are constrained by the following formula to obtain the Weibull parameters:

[0280] λ ∈ (0, +∞)

[0281] k ∈ (1, +∞)

[0282] Among them, k = 1 is the constant failure rate for random failures, and k > 1 is the increasing failure rate for aging failures.

[0283] Furthermore, the system further includes a 3D modeling module for:

[0284] Match and fuse the 3D BIM model coordinate system of the actuator with the actual geographic coordinate system to obtain a coordinate matching model;

[0285] Assign a unique ID to each component in the coordinate matching model, where the ID corresponds one-to-one with the actuator, to obtain an ID mapping relationship table;

[0286] Based on the ID mapping relationship table, associate the actuator operation data and the fault prediction probability with the corresponding components of the 3D BIM model to obtain a 3D model with dynamic data binding;

[0287] Send the 3D model with dynamic data binding to the terminal; the 3D model with dynamic data binding is used for display on the terminal display device;

[0288] Among them, the terminal is used to feedback virtual instructions; the virtual instructions are used to instruct the actuator to change its own operating state.

[0289] Furthermore, after sending the 3D model with dynamic data binding to the terminal, it further includes:

[0290] Receive virtual device instructions; the virtual device instructions are feedback by the terminal;

[0291] Generate a virtual device based on the virtual device instructions and the BIM model;

[0292] Connect the virtual device to the 3D model with dynamic data binding to obtain a new 3D model;

[0293] Execute the control strategy, drive the new 3D model to act, and calculate the dynamic response of the new 3D model;

[0294] Based on the dynamic response, verify the logic compliance and generate a virtual device result prediction;

[0295] Generate a virtual commissioning report based on the virtual device result prediction.

[0296] Furthermore, after sending the 3D model with dynamic data binding to the terminal, it further includes:

[0297] Receive preset scenario instructions; the preset scenario instructions are feedback by the terminal;

[0298] Inject faults into the 3D model with dynamic data binding based on the predicted scenario to obtain a 3D model with faults;

[0299] Based on the 3D model with faults, predict the fault impact through the AI model to generate a fault impact prediction;

[0300] Generate fault handling steps based on the fault impact prediction;

[0301] Generate an optimal emergency plan based on the fault handling steps.

[0302] Furthermore, the system also includes a 5G transmission module for:

[0303] When it is detected that the operation data of the actuator has not been received through the fieldbus for more than 200 ms, generate a bus anomaly signal;

[0304] Based on the bus anomaly signal, asymmetrically encrypt the executable policy using the public key of the actuator to obtain the encrypted policy;

[0305] Calculate the hash of the encrypted policy to obtain the encrypted policy hash value;

[0306] Broadcast the encrypted policy hash value and the encrypted policy to the blockchain network to obtain the blockchain transaction hash;

[0307] Send the encrypted policy and the blockchain transaction hash to the target actuator through the 5G network, for instructing the target actuator to feedback the hash verification result through the 5G network;

[0308] Among them, the blockchain network permanently stores the encrypted policy hash value and the encrypted policy.

[0309] Furthermore, the optimization parameter module is also used for:

[0310] Normalize the electricity price curve, load demand and actuator operation data to obtain normalized data;

[0311] Encode the normalized data into a state vector;

[0312] Input the state vector into the deep deterministic policy gradient model to output a continuous action vector;

[0313] Perform constraint processing on the continuous action vector to obtain optimization parameters;

[0314] Among them, the deep deterministic policy gradient model calculates the continuous action vector through the following formula:

[0315]

[0316] Q(s t ,a t )=r t +γ·Q'(s t+1 ,μ'(s t+1 ))

[0317] Among them, is the gradient of the policy objective function with respect to the policy network parameters, s i is the state of the i-th sample, μ(s i) is the action generated by the policy network according to the state, and Q(s t , a t ) is the Q-value prediction of the current state, r t is the immediate reward at the current moment, and γ is the discount factor;

[0318] Among them, the model is updated through the following formula:

[0319] θ Q' ← τθ Q + (1 - τ)θ Q’ , θ μ' ← τθ μ + (1 - τ)θ μ’

[0320] Among them, τ is the update rate.

[0321] For the system embodiment, since it basically corresponds to the method embodiment, the relevant parts can be referred to the partial description of the method embodiment. The system embodiments described above are only illustrative. The components described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the present disclosure solution. Those of ordinary skill in the art can understand and implement it without creative efforts.

[0322] The above embodiments only represent several implementation manners of the embodiments of the present application. The description is relatively specific and detailed, but it should not be construed as a limitation on the patent scope of the embodiments of the application. It should be noted that for those of ordinary skill in the art, without departing from the concept of the embodiments of the present application, several modifications and improvements can still be made, and these all belong to the protection scope of the embodiments of the present application.

Claims

1. A redundant control method for a gas power plant based on multimodal perception and AI decision-making, characterized in that: The method comprises: receiving the operating data of the actuator; the operating data of the actuator is multimodal data collected by the sensor of the actuator and pre-processed by the edge gateway; the actuator is a device in the gas power plant responsible for executing instructions; Calculating the failure prediction probability of the actuator through an LSTM neural network according to the operating data of the actuator; Calculating the optimization parameters of the actuator by using a reinforcement learning algorithm according to the operating data of the actuator; Based on the fault prediction probability and the optimization parameter, generating an executable strategy; The executable strategy is sent to the execution mechanism; the executable strategy is used to instruct the execution mechanism to adjust its own operating parameters.

2. According to claim 1, the redundant control method of a gas power plant based on multimodal perception and AI decision-making is characterized in that: The method of calculating the fault prediction probability of the actuator through an LSTM neural network based on the operating data of the actuator includes: Inputting the actuator operation data into a bidirectional LSTM layer to extract time series features and obtain a feature vector; The bidirectional LSTM layer extracts time series features using the following formula: in, is the forward hidden state, is the backward hidden state, g(·) is the fusion function; Performing sub-window segmentation on the feature vector to obtain a window feature vector; Inputting the window feature vector into the convolution layer for convolution to obtain a feature matrix after spatiotemporal focusing; Dynamically weighting key features of the feature matrix to obtain a weighted feature matrix; Inputting the weighted feature matrix into the prediction layer to obtain a prediction result; The prediction results are integrated to obtain the failure prediction probability of the actuator.

3. The redundant control method for a gas power plant based on multimodal perception and AI decision-making according to claim 2 is characterized in that: The prediction layer obtains the prediction result by the following method: Inputting the weighted feature matrix into the global average pooling layer to obtain a pooled feature matrix; The pooled feature matrix is ​​reduced in dimension through a fully connected layer to obtain reduced-dimensional features; Activating the dimension-reduced features to obtain a fault probability scalar; the fault probability scalar represents the fault probability of each of the actuators; Inputting the weighted feature matrix into the GRU layer to extract mid-term time series features to obtain mid-term features; Input the extracted features into the time attention layer to obtain time window features; and performing linear regression on the time window characteristics to obtain the daily performance degradation rate of the actuator; Inputting the weighted characteristic matrix into the Weibull model to obtain Weibull parameters; Inputting the Weibull parameter into the life calculation layer to obtain the remaining life of the actuator; Among them, the loss function is calculated by the following formula: in, is the improved Focal Loss loss function, is the Huber Loss function, is the Weibull negative log-likelihood; The weight is calculated by the following formula: Among them, σ i is a trainable parameter; Back-propagating the loss function to update the prediction layer; A prediction result is obtained based on the failure probability scalar, the daily performance degradation rate and the remaining life.

4. The redundant control method for a gas power plant based on multimodal perception and AI decision-making according to claim 3 is characterized in that: The Weibull model obtains Weibull parameters by the following method: Splitting the weighted feature matrix into multiple heads to obtain a multi-head feature matrix; Calculate the QKV matrix independently for each of the multi-head feature matrices to obtain an attention subspace; Fusing the attention subspace features to obtain a long-term feature matrix; Based on the long-term characteristic matrix, the original Weibull parameter is calculated by the following formula: λ,k=W λ ·h attn +b λ ,W k ·h attn +b k Among them, λ is the scale parameter, which controls the change of life, and k is the shape parameter, which controls the trend of failure rate change; The original Weibull parameter is constrained by the following formula to obtain the Weibull parameter: λ∈(0,+∞) k∈(1，+∞) Among them, k=1 is a constant failure rate, which is used for random failures, and k>1 is an increasing failure rate, which is used for aging failures.

5. The redundant control method for a gas power plant based on multimodal perception and AI decision-making according to claim 1 is characterized in that: After receiving the operating data of the actuator, the method further includes: Matching and fusing the three-dimensional BIM model coordinate system of the actuator with the actual geographic coordinate system to obtain a coordinate matching model; Assigning a unique ID to each component in the coordinate matching model, wherein the ID corresponds to the actuator in a one-to-one manner, and obtaining an ID mapping relationship table; Based on the ID mapping relationship table, the actuator operation data and the fault prediction probability are associated with corresponding components of the three-dimensional BIM model to obtain a three-dimensional model with dynamic data binding; The three-dimensional model bound with the dynamic data is sent to a terminal; the three-dimensional model bound with the dynamic data is used to be displayed on a display device of the terminal; The terminal is used to feed back virtual instructions; the virtual instructions are used to instruct the actuator to change its own operating state.

6. The redundant control method for a gas power plant based on multimodal perception and AI decision-making according to claim 5 is characterized in that: After sending the three-dimensional model bound with the dynamic data to the terminal, the method further includes: receiving a virtual device instruction; the virtual device instruction is fed back by the terminal; Generate a virtual device based on the virtual device instructions and the BIM model; Connecting the virtual device to the dynamic data-bound three-dimensional model to obtain a new three-dimensional model; Executing a control strategy to drive the new three-dimensional model to move and calculate a dynamic response of the new three-dimensional model; generating a virtual device result prediction based on the dynamic response verification logic compliance; Based on the virtual device result prediction, a virtual commissioning report is generated.

7. The redundant control method for a gas power plant based on multimodal perception and AI decision-making according to claim 5 is characterized in that: After sending the three-dimensional model bound with the dynamic data to the terminal, the method further includes: Receiving a preset scene instruction; the preset scene instruction is fed back by the terminal; Performing fault injection on the three-dimensional model bound with dynamic data based on the prediction scenario to obtain a three-dimensional model of the fault; Based on the three-dimensional model of the fault, predict the impact of the fault through the AI ​​model to generate a fault impact prediction; generating fault handling steps based on the fault impact prediction; Based on the fault handling steps, an optimal emergency plan is generated.

8. The redundant control method for a gas power plant based on multimodal perception and AI decision-making according to claim 1 is characterized in that: The method further comprises: When it is detected that the operating data of the actuator is not received via the field bus for more than 200ms, a bus abnormality signal is generated; Based on the bus abnormality signal, the execution strategy is asymmetrically encrypted using the public key of the execution mechanism to obtain an encrypted strategy; Calculating a hash for the encrypted policy to obtain a hash value of the encrypted policy; Broadcasting the encrypted policy hash value and the encrypted policy to the blockchain network to obtain a blockchain transaction hash; Sending the encrypted policy and the blockchain transaction hash to the target execution agency through the 5G network to instruct the target execution agency to feedback the hash verification result through the 5G network; Wherein, the blockchain network permanently stores the encrypted policy hash value and the encrypted policy.

9. The redundant control method for a gas power plant based on multimodal perception and AI decision-making according to any one of claims 1 to 8, characterized in that: The step of calculating the optimization parameters by using a reinforcement learning algorithm based on the operating data of the actuator includes: Normalizing the electricity price curve, load demand and the operating data of the actuator to obtain normalized data; encoding the normalized data into a state vector; Inputting the state vector into a deep deterministic policy gradient model and outputting a continuous action vector; Performing constraint processing on the continuous motion vector to obtain the optimization parameter; Among them, the deep deterministic policy gradient model calculates the continuous action vector through the following formula: Q(s t ,a t )=r t +γ·Q'(s t+1 ,μ'(s t+1 )) in, is the gradient of the policy objective function to the policy network parameters, s i is the state of the i-th sample, μ(s i ) is the action generated by the policy network according to the state, Q(s t ,a t ) is the Q value prediction of the current state, r t is the instant reward at the current moment, γ is the discount factor; Among them, the model is updated by the following formula: i Q' ←tth Q +(1-τ)θ Q′ ,i μ' ←tth μ +(1-τ)θ μ′ Among them, τ is the update rate.

10. A redundant control system for a gas power plant based on multimodal perception and AI decision-making, characterized in that: The system comprises: A receiving module, used for receiving the operating data of the actuator; the operating data of the actuator is multimodal data collected by the sensor of the actuator and pre-processed by the edge gateway; the actuator is a device in the gas power plant responsible for executing instructions; A fault prediction module, used to calculate the fault prediction probability of the actuator through an LSTM neural network according to the operating data of the actuator; An optimization parameter module, used to calculate the optimization parameters of the actuator through a reinforcement learning algorithm according to the operating data of the actuator; A decision module, used for generating an executable strategy based on the fault prediction probability and the optimization parameter; The adjustment module is used to send the executable strategy to the execution mechanism; the executable strategy is used to instruct the execution mechanism to adjust its own operating parameters.

Images